Mist is an image preprocessing tool designed to protect artwork from AI-driven style mimicry by applications like Stable Diffusion and NovelAI. It adds imperceptible watermarks that render images unrecognizable to these models, preventing effective style replication and producing scrambled outputs. The project is suitable for artists and content creators seeking to safeguard their digital creations.

How It Works

Mist employs an adversarial perturbation approach, modifying images to be unrecognizable by diffusion models. It integrates with Stable Diffusion, requiring an open gradient flow for its end-to-end adversarial perturbation process. This involves modifying specific Stable Diffusion model files to enable the watermark injection.

Quick Start & Requirements

• Install: conda env create -f environments.yml, conda activate mist, pip install --force-reinstall pillow.
• Prerequisites: Stable Diffusion v1.4 checkpoint (sd-v1-4.ckpt), advertorch 0.2.4.
• Setup: Requires environment setup via environments.yml and downloading the Stable Diffusion checkpoint.
• Docs: Documentation, Homepage

Highlighted Details

• Watermarking strength, steps, output size, blocking, and mode are configurable.
• Includes scripts for evaluating robustness against cropping and resizing.
• Paper accepted for Oral presentation at ICML 2023.

Maintenance & Community

• The project is associated with the psyker-team.
• Links to further documentation and a homepage are provided.

Licensing & Compatibility

• Licensed under GPL-3.0.
• Requires modifications to Stable Diffusion files for gradient flow, potentially impacting compatibility with unmodified Stable Diffusion codebases.

Limitations & Caveats

The project requires specific versions of dependencies and modifications to Stable Diffusion's internal files, which may complicate integration and maintenance. The GPL-3.0 license may impose copyleft restrictions on derivative works.